Skip to content
rinselabs

Inside the lab

Ideas worth investigating.
Systems worth building.

Our product direction and conceptual research themes. Research directions describe questions and possible methods — not completed experiments, working prototypes, or published results.

Conceptual interconnected processing architecture representing agent orchestration
Product developmentIn Development

Rinse Labs AI Agent Platform

Problem
Teams experimenting with agents lack a single place to build, connect, orchestrate, and govern them.
Approach
A platform for agent creation, workflow orchestration, tool connections, approvals, and execution observability.
  • Agent orchestration
  • Tool APIs
  • Human-in-the-loop
See the platform vision
Conceptual compact model lattice held between precision mesh layers
Model experimentsResearch Direction

Small-model specialization study

Problem
When does a small, adapted language model outperform a frontier model on a narrow business task?
Approach
Fine-tuning and quantization compared against prompted baselines on domain evaluation sets.
  • SLMs
  • Fine-tuning
  • Evaluation

Conceptual exploration — not completed research

Conceptual evaluation gates checking an output pathway
Evaluation & optimization toolsResearch Direction

Agent reliability harness

Problem
Agent behaviour regresses silently when prompts, models, or tools change.
Approach
An evaluation runner that replays golden task sets and scores completion, tool-call accuracy, and cost.
  • Eval harness
  • Tracing
  • CI gates

Conceptual exploration — not completed research

Connected agent processors sharing an orchestration hub
Agentic systemsResearch Direction

Multi-agent research pipeline

Problem
Research tasks require gathering, cross-checking, and synthesizing many sources with citations.
Approach
Planner, researcher, and validator agents with human review before output is accepted.
  • Multi-agent
  • RAG
  • Tool use

Conceptual exploration — not completed research

A reward path navigating an experimental grid environment
Experimental researchResearch Direction

RL for agent behaviour

Problem
Can reward-shaped training make multi-step agent behaviour more consistent than prompting alone?
Approach
Task-level reward signals compared against prompted baselines on completion and error rates.
  • Reinforcement learning
  • Agent eval

Conceptual exploration — not completed research

Modular engineering architecture with verification connections
AI developer toolsResearch Direction

AI pair-engineering workflow

Problem
AI coding tools produce volume fast, but review and verification become the bottleneck.
Approach
Claude/Codex-assisted implementation wrapped in spec-first planning, generated tests, and structured review.
  • Claude
  • OpenAI Codex
  • Test generation

Conceptual exploration — not completed research

The flagship: an AI Agent Platform.

In development
Concept preview — platform in development
Workflow builder · contract-review
needs revision ↶passRequestevent / incomingPlanneragent / runningResearchdocs + memoryExtracttools / scoped APIEvaluaterules + evidenceApprovalhuman / waitingMODEL · KNOWLEDGE · TOOL CONNECTIONS
Execution history
  • Contract reviewawaiting approval
    run_8f2a · illustrative
  • Support triagecompleted
    run_8f29 · illustrative
  • Weekly reportcompleted
    run_8f27 · illustrative
Approval queue

Update renewal terms in CRM

Requested by QA Validator · 3 sources cited

Monitoring (sample data)

Illustrative values only — no live data.

Concept preview — platform in development. Not a live application.

Want to build alongside us?

We take on applied research and engineering collaborations where the problem is genuinely hard.

Work With Us About the platform